Control of Arterial Hemoglobin Saturation in Premature Infants using H-infinity Synthesis and Performance Specifications from Best Clinical Practice

Transcription

1 Control of Arterial Hemoglobin Saturation in Premature Infants using H-infinity Synthesis and Performance Specifications from Best Clinical Practice A Thesis presented to the Faculty of the Graduate School University of Missouri Columbia In Partial Fulfillment of the Requirements of the Degree Master of Science By DANIEL QUIGLEY Dr. Roger Fales, Advisor July 2013

2 The undersigned, appointed by the Dean of the Graduate School, have examined the thesis entitled CONTROL OF ARTERIAL HEMOGLOBIN SATURATION IN PREMATURE INFANTS USING H-INFINITY SYNTHESIS AND PERFORMANCE SPECIFICATIONS FROM BEST CLINICAL PRACTICE Presented by Daniel Quigley A candidate for the degree of Master of Science And hereby certify that in their opinion it is worthy of acceptance. Professor Roger Fales Professor Craig Kluever Professor Carmen Chicone

3 Acknowledgments I would like to thank my supervisor, Dr. Roger Fales, whose technical expertise and experience with the subject was priceless in guiding me through the research process. Also for his flexibility, patience, and genuine caring attitude that made this thesis project personal and fun. I would like to thank my family, friends, and teammates who have always helped push me to pursue excellence in everything I do. A special thanks to Shannon Leinert for being my support team as we worked together to finish our degrees on time. Finally I would like to thank my committee members Dr. Craig Kluever and Dr. Carmen Chicone for their time. ii

4 TABLE OF CONTENTS Acknowledgments... ii List of Figures... iv List of Tables...v ABSTRACT... vi Chapter 1: Introduction Background Literature Review Theoretical work Prior Clinical Testing of similar devices Prior work at the University of Missouri Objectives and motivation Chapter 2: Model Considered Models Three Transfer function model Genetic Algorithm Non-linear model Chapter 3: Control design Performance Specifications H-infinity synthesis Performance Analysis Chapter 4: Results Controller Synthesis results Open loop parameter estimation Constant parameter estimation Changing parameter estimation Parameter estimation with real data Closed loop simulations Chapter 5: Prototype design and testing Device Performance Specifications Drawings and Final Design Chapter 6: Discussion and Conclusion Discussion Open loop controller selection Closed loop controller selection Recommendations Conclusion Appendix A: Model information...72 Appendix B: Controller Results...73 Appendix C: Prototype Drawings...75 Appendix D: Genetic Algorithm MATLAB code...79 iii

5 References...84 List of Figures Fig. 1 Data collection setup [23] Fig. 2 Three transfer function model of infant response Fig. 3. Multiplicative uncertainty for nominal plant Fig. 4. Model block diagram with multiplicative uncertainty Fig. 5. GA window example with estimated model shown to match the feedback information Fig. 6. Flowchart of the parameter estimating genetic algorithm model Fig. 7. Parameter estimation simulations with closed loop FiO 2 control Fig. 8. Block diagram of non-linear model of FiO 2 SpO 2 relationship Fig. 9 Block Diagram of model, performance weights, and uncertainty model Fig. 10. Closed loop step response of controller without rate limiting performance specification Fig. 11. Response of a controller synthesized using the rate limiting specification Fig. 12. Controller 1 parameter estimation simulation Fig. 13. Error used to assign fitness of parameter sets during simulations Fig. 14. Yellow line represents the correct choice, red is simulation result Fig. 15. Result when the correct parameters are between controllers Fig. 16. Estimating varying time constants test Fig. 17. Controller selection for varying time constant test Fig. 18. Estimating varying time constants test Fig. 19. Controller selection, varying parameters test Fig. 20. Changing gain parameter estimation Fig. 21 Changing gain controller selection Fig. 22. Raw data simulation with unknown model parameters Fig. 23. Raw data simulation controller selection Fig. 24. Input data from clinical setting used in GA testing Fig. 25. Closed loop parameter estimation Fig. 26. Closed loop control test SpO 2 and control effort Fig. 27. Modeled closed loop results with real disturbance inputs from data Fig. 28. Real disturbance input data used Fig. 29. NEO2 Blend valve retrofit design assembly drawing Fig. 30. Control panel and hardware box Fig. 31. Original prototype design, circa Fig. 32. New prototype design completed Fig. 33. Bode magnitude response of controller K for nominal plant iv

6 List of Tables Table 1 Parameter uncertainty ranges for FiO 2 transfer function Table 2 Parameter ranges for Heart Rate and Respiratory Rate Table 3 Performance and Stability verification values Table 4 Uncertainty Model (W A ) for each of the six plants Table 5 Uncertainty Model for Heart Rate and Respiratory Rate Table 6 Results of controller synthesis v

7 ABSTRACT This work details the development of a novel control system for regulating the Arterial Hemoglobin saturation (SpO 2 ) of premature infants. An H-infinity synthesis method was used to derive multiple controllers to cover a large range of parametric uncertainty in the model. Performance specifications were defined using manual control protocols, which are clinically proven to reduce the rate of Retinopathy of Prematurity (ROP) in the target patients. The controllers were tested using previously validated models and shown to meet the robust stability and performance specifications. In order to switch between controllers and adapt to the changing physiology of the patients a genetic algorithm was modified and tested. This algorithm was proven to adequately track changing model parameters and select an appropriate robust controller for providing closed loop control. In addition a physical prototype was designed and built to facilitate automatic control of FiO 2 during clinical tests of the control algorithms. vi

8 Chapter 1: Introduction 1.1 Background Respiratory distress syndrome (RDS) is a breathing disorder in preterm infants, which is caused by underdeveloped lungs and insufficient surfactant production[1].this syndrome is primarily seen in premature infants with the youngest gestational age being the most susceptible but a very small percentage suffer from a genetic inability to produce surfactant. The lack of surfactant in the alveoli inhibits the lung s ability to properly inflate, which is necessary for ventilation. RDS is the leading cause of death for infants less than 31 weeks gestational age[2]. Oxygen enriched air, with the level of oxygen described as the fraction of inspired oxygen (FiO 2 ), is supplied to the infant and has been a standard protocol to combat RDS since the early 1940 s[3]. This oxygen therapy is used with respiratory support devices like a nasal cannula which is a simple tube that delivers oxygen to the nostrils, ventilator which mechanically moves air in and out of the lungs, CPAP which provides continuous positive airway pressure but the patient breaths on their own, or other respiratory support modalities. Despite the use of artificial surfactant therapy, the need for supplemental oxygen persists. There are many health risks associated with supplemental oxygen, however. High blood oxygen concentrations can cause retinopathy of prematurity (ROP). ROP is a disease of the eye in which the vascularity of the eyeball grows in a rapid and disorganized manner. Preterm infants are susceptible because the eyeball develops late in the gestation period. The rapid growth of the retinal blood vessels is in part incited by episodes of hyperoxia, which can result in retinal scaring and possible blindness. This is 1

9 an obvious drawback to supplemental oxygen therapy for premature infants. While some cases of ROP naturally correct themselves, many need surgery to regain sight for the infant. In extreme cases total blindness can occur such as the case of famous musician and entertainer Stevie Wonder[4]. The risk of ROP is associated with high arterial hemoglobin saturation (SpO 2 ), but low SpO 2 concentrations can result in hypoxia and tissue damage[5]. For these reasons it is very important to maintain a safe SpO 2 range for these at risk patients. This is done in hospitals across the world by nurses and doctors using a manually operated analog gas mixing device. Typically digital sensors provide feedback on the FiO 2, SpO 2, heart rate (HR), and respiratory rate (RR). The level of FiO 2 supplied to the infant is based on nurses judgment and protocols for adjusting it are typically mandated by the individual medical doctors at the hospital. Arterial oxygen saturation is most commonly monitored by pulse oximetery. This is a method which improved upon the more invasive indwelling umbilical catheter electrode by using a laser to determine the SpO 2. The laser detects the red color of the blood while flowing through extremities and determines the saturation of oxygen by the richness of this red color. The current techniques used allow for fast sampling and averaging algorithms have been able to give accuracy in the range of ±2% for saturations above 70% [6]. The desired SpO 2 range as agreed upon by the literature has fluctuated since the early studies but the most current research shows 85% to 93% is the most effective range for avoiding ROP and hypoxia [7]. There are alarms present in the neonatal intensive care unit (NICU) that notify medical attendants of a dangerous SpO 2. However, it has 2

10 been shown that preterm infants spend approximately half of the time was spent within the clinically intended range of SpO 2. More specifically 20% of the time was spend below the range and30% of the time above this range during routine neonatal intensive care [8, 9]. One study even revealed that the upper limit alarms had been set too high 76% of the time because of the frequency that they were exceeded[10]. These statistics indicate that there is still a large potential for improvement of the time spent in a healthy range despite advancements in technology and knowledge of the subject. A small number of researchers have been working on using closed loop automatic control techniques to do this since the late 1970 s; Beddis et al, Dugdale et al, Bhutani et al, Morozoff et al, Claur et al, and Urschitz et al. An overview of these works with their significance is provided. One of the most challenging aspects of applying automatic control is the changing physiology of the patients. There is a wide range of gestational ages, birth weights, and varying degrees of health that affect the response of each infant to FiO2 adjustments. In addition, as each infant grows the changing physiology represents the need for a controller that can recognize and adapt to the varying parameters or one that is robust enough to provide effective control for all parameters. The literature shows researchers have chosen one these two options but there is potential for a robust controller that is also adaptive. 3

11 1.2 Literature Review Theoretical work Several control systems for the ventilation of premature infants have been presented in literature. Tehrani et al. proposed using a proportional, integral, derivative (PID) controller with feedback arterial partial pressure of oxygen in newborn infants in 1991 [11]. The performance of the control system was investigated using computer simulation. The results under two different test conditions indicate stable performance for each case. Many of the other controllers in literature include adaptive components. Yu et al. developed a multiple-model adaptive controller (MMAC) to regulate oxygen saturation by changing FiO 2 in 1987 [12]. The MMAC procedure assumes that the system can be represented by one of a finite number of models. A controller is designed for each of these models that have constant parameters. The actual control output is a weighted sum of each individual controller. The MMAC was found to effectively regulate oxygen saturation in animal studies. Taube et al. developed an adaptive PID control system based on the model by Grodins that was then simplified by Sano and Kikucki to a static relationship in 1991 [9, 15, 16]. The adaptive part of the controller involves updating the KL gain by a ratio of the partial pressure of arterial oxygen and the input FiO 2. They found that the automatic regulation of FiO 2 may be a better method than manual adjustments. 4

12 1.2.2 Prior Clinical Testing of similar devices A history of past clinical trials was used to more fully understand the progression of the available computing technology, control algorithm research, and understanding of the patient s needs. A brief synopsis of these clinical trials follows. The first published study of closed-loop automatic control of inspired oxygen was done in 1979 using arterial oxygen tension (PaO 2 ) feedback on a 1 minute interval from an indwelling umbilical electrode[13]. The FiO 2 was controlled in discrete steps of 5% and used with intermittent positive pressure ventilation, continuous positive airway pressure, and hood support. Limits for PaO 2 were set between 7.3 kpa and 10.7 kpa. The amount of time spent outside these limits decreased from 27.6% to 12.2% with use of servo control compared to manual. The mean duration of the episodes below the limits decreased from 6.1 to 2.5 minutes and mean duration above decreased from 4.7 to 1.7 minutes. However, the number of episodes above the set limits actually increased during servo control. This could have been an issue with the sensor feedback or a problem inherent in the control system, which indicates there is room for improvement of either control algorithm or equipment capability. The second clinical study involving infants was done in 1985 and also utilized an indwelling umbilical electrode[14]. The target value for PaO 2 was chosen to be 10 kpa measured every minute, and FiO 2 was controlled in steps of 5 or 10%. The authors considered adaptive control options for this application, but decided on a simple controller based on an estimated steady state gain and time constant of the response to a step input. This controller was able to tolerate up to 300% variation in static gain parameters. The researchers chose to avoid adaptive control because non-linear problems 5

13 are difficult to analyze and adaptive controllers are more complicated than conventional options. Control parameters were chosen conservatively (1 kpa per % O 2 gain and a 120 second time constant) to ensure stability over a wide range of operating conditions because quality of control was only a secondary goal. The mean time spent within 1kPa of the set point was 74.9% with automatic control, and 45.2% without automatic control. The authors proposed that an adaptive controller could be developed in the future with the goal of improving care. At the time of the study, a parameter-estimating algorithm was suggested, but the required level of engineering sophistication for this was not available. By 1993 pulse oximeter technology had developed enough that measuring the blood oxygen saturation (SaO 2 ) with non-invasive devices was possible. The first automatic control study to be done using this technology was published by Morozoff et al. They used the feedback to define discrete states that described the direction of the error, velocity, and acceleration while ignoring the magnitude of those values. The states merely describe whether the SaO 2 is above or below the setpoint, and if the SaO 2 is trending toward or away from the set point. FiO 2 is then either increased or decreased relative to its previous value. This controller required very little processing power and was adept at controlling the SaO 2 except when helping recover from sudden desaturation events. In these instances manual intervention was required. The results showed that the acceptable range of 90-95% SaO 2 was maintained 50% of the time during automatic control while only 39% of the time under manual control[15]. A study done in 2001 is the first to use a novel closed loop controller in comparison to a fully dedicated nurse performing continuous monitoring and adjustments 6

14 as opposed to routine monitoring [16]. This study used SpO 2 feedback and a set of rules to adjust the FiO 2 when outside the range of 88-96%. These rules covered cases including slow changes during apparent stability and rapid cases of desaturation. The mean weight and gestational age was 712g and 25 weeks. The SpO 2 feedback period was chosen to be 2 seconds but FiO 2 adjustments were allowed to be continuous. Analysis showed that manual and automatic control resulted in similar frequency and mean duration periods of hypoxemia, but under automatic control the number of overshoots to hyperoxemic ranges was reduced from 5.8±7.8% to 1.5±1.9%. The percent of time spent in the goal range of 88%-96% was improved from 66.3±13.9% to 74.9±12.5%. This study is the first to truly prove the efficacy of closed loop control as a replacement for a fully dedicated nurse or regularly attending nurse. This control system doesn t require manual intervention as did the controller in the study by Morozoff and Evans, and proves a savings in nursing time since normal conditions require nurses to perform other tasks in addition to constantly monitoring SpO 2. Urschitz et al. furthered the research in 2004 with another clinical trial testing a new style of controller. This controller is based on artificial intelligence research and a method called temporal-abstraction, which helps filter the feedback and reduce the number of adjustments the controller must make [17]. The controller was programmed by monitoring clinician FiO 2 adjustments but, similarly to previous work, was not able to respond to acute severe hypoxic episodes since it had a built-in wait period of 180 seconds after each adjustment. It was, however, able to increase the amount of time SpO 2 was maintained in an acceptable range by 10-15% [18]. Urschitz el al. also made reference to the need for more research to determine an optimal range for SpO 2, as well 7

15 as the need for further work on improving the sensor technologies to avoid false feedback. The team of Morozoff and Smyth published results testing three different controller options, including the first adaptive controller, in The first controller was called a State Machine Algorithm that uses the error, error velocity and error acceleration as inputs to a set of rules that determine the FiO 2 change similar to their previous controller tested in The second controller tested was a classic PID style controller with gains that could be changed on the fly by user input. The adaptive controller uses a linearized model of the oxygen dissociation curve and either increases or decreases the slope of certain linear segments as dictated by the SpO 2 feedback [19].The adaptive model proved best and had the smallest need for manual intervention; in second was the state machine, and finally the PID. All three controllers improved upon manual control[20]. The significant advancement from this work is the effective use of an adaptive controller that is capable of changing parameters with changing neonate physiology. Claure et al. also pursued new research in this area and published another clinical study in They used an oximeter with an eight second feedback interval and allowed the control system to adjust FiO 2 every second. The specifics on the controller used were not published, but it considers an averaged basal FiO 2 as a moving baseline and makes adjustments based on that when the SpO 2 is out of the intended range. A safety measure that was introduced would select the median FiO 2 from the past 60 seconds if the feedback signal was lost for more than 10 seconds. They were able to show that the automatic control increased the time the infants spent in the range of 88-8

16 95% from 42%±9% to 58%±10%[21]. It also showed an increased time the infants spent at lower FiO 2 settings, which is thought to help avoid ROP. There were, however, an increased number of episodes below the desired range, which do not correspond with oscillations or hyperoxemic events, suggesting that fine-tuning of the controller is possible. Claure et al. continued their work with a much larger study in 2011 with 32 infants at four different hospitals using the same automatic controller as before. They recorded two consecutive 24-hours periods, one with manual FiO 2 control and the other with automatic control. This time a target range of 87%-93% was used and again improvements over manual control were seen. Also, the significant increase in time spent below the desired range was seen as before[22]. The authors note that the infants chosen to participate in the study were those with the most frequent fluctuations in SpO 2 and therefore they possessed the greatest risk for ophthalmic or neurologic injury Prior work at the University of Missouri Two previous graduate students at the University of Missouri have conducted prior non-clinical published investigations of the device. Bradley Krone, a graduate student who has since received his M.S. for his work on this project, analyzed the relationship between FiO 2 and SpO 2 and found that it can be modeled using a first order transfer function. Krone concluded that through the analysis of the biological system and recorded data it is thought that by considering the relationship between the SpO 2 and the FiO 2, HR, and RR that a model will be able to better simulate the SpO 2 response to inputs [23]. Several different transfer function models were tested, and the dynamic 9

17 transfer function model was found to model the SpO 2 accurately for a prolonged amount of time. On average, the function is able to model SpO 2 adequately for 15 minutes or more. Furthermore, the model is reliable over many situations and can adequately estimate the gains and time constants in simulated data [12]. Thus, the controller operates using a model that accurately and reliably simulates real-life data. Next, tests were performed in order to determine the best controller to use. The test compared the parameters of maximum SpO 2, maximum FiO 2, average SpO 2, and average FiO 2. Three different controller types were tested for a simulation with constantly changing FiO 2, HR and RR to better simulate the changing parameters of actual infants. These controllers included a linear quadratic regulator proportional integral controller (LQR-PI), a robust controller designed using H-infinity optimization, and an adaptive controller with feed forward disturbance rejection (AC-FFDR) [23]. The results showed that all controllers attempted to reject the disturbances caused by variations in HR and RR and keep the SpO 2 at a given set point, and that the robust controller has the best performance because it has the lowest maximum SpO 2 and average SpO 2. While the robust controller had the largest maximum FiO 2 and average FiO 2 values, these are well within the limitations of the FiO 2 [23]. Keim, a former student at the University of Missouri who has earned his Ph.D. for his work on this project, further addressed the effectiveness of the device controls in his dissertation Control of Arterial Oxygen Saturation in Premature Infants [24]. To further investigate the controller s performance, actual infant response data was collected during a study performed at Columbia Regional Hospital, now known as University Women s and Children s Hospital. The patients involved were premature infants with frequent 10

18 desaturation periods. Clinical data was collected for patients on ventilators, FlowPAP, and SiPAP respiratory support devices which are both specialized devices for positive airway pressure. A LabView program collected heart rate, respiratory rate, SpO 2 and FiO 2 signals. The oxygen sensor transmitted data every second, and the heart rate, respiratory rate, and SpO 2 were collected every 5 seconds. The pulse oximeter recorded these values. The data collection program runs continuously saving one-hour long files of the signals. To more completely capture the effects of surrounding circumstances, nurses kept diary files of the patient care-giving information. Such information included gestation age, adjusted age, respiratory modality, oxygen alarm parameters, and current medications. Three different estimation systems were evaluated, including a dynamic fuzzy logic system, a continuous parameter-estimating extended Kalman filter, and a discrete parameter-estimating extended Kalman filter [24]. An adaptive controller was developed based on the discrete extended state observer and parameter-estimating extended Kalman filter system. The adaptive controller was found to display the best performance for a system with unknown disturbances and large ranges of model parameters when compared to a static proportional integral and a robust controller. It displayed the fastest and most accurate performance in the disturbance rejection simulation. First, the microcontroller receives the patient s SpO 2, heart rate, respiratory rate, and percentage of oxygen through serial cables. The adaptive control system then uses these signals to determine the optimal oxygen percentage to administer to the infant. It will adjust FiO 2 levels and also isolate the apnea alarm sent from the bedside monitor, using it to activate the flow control motor and vibration motor [24]. Keim concludes that the software is shown to run in real 11

19 time in less than 5 seconds. He continues saying that this prototype should be capable of administering the same level of patient care as a nurse. In addition to being at least as good as a nurse, the prototype should aid the nurse in providing a more consistent and higher level of care for the premature infants in the NICU, [24]. These two researchers, Krone and Keim, based their work on data they collected in a clinical setting from target patients. The hospital sensors were used to collect feedback information on a laptop. A representation of the sensor setup and data collection procedure can be seen in Fig. 1. The data they collected was analyzed and used to justify the models they created. This data was also used to verify the model and controller that is presented in this work. Fig. 1 Data collection setup[23] 1.3 Objectives and motivation As discussed above a recent study testing strict guidelines for increasing and decreasing, or weaning FiO 2 and monitoring O 2 saturation manually, the researchers showed a significant reduction in rates of ROP [1]. Those results were repeated in a larger study and helped prove that proper manual control can be very effective in reducing the incidence of ROP [25]. As these advances in knowledge about the physiology and health of the patients continue to improve there are new opportunities for 12

20 closed loop control to match the success of the medical research. Better understanding of the patient s needs can be coupled with improved computing options and control theory to achieve new standards of care. Multiple clinical studies have proven the efficacy of an automatic control algorithm for managing the FiO 2 compared to a nurse s control [13-16, 18, 20-22]. A few of these studies have shown a significant decrease in ROP rates and represent a reduced workload for the attending nurse if employed full time. Many different control strategies have been applied from simple proportional integral derivative control to fuzzy logic, a state machine rule set, and even an adaptive model based on the O 2 dissociation curve [19]. There is a noticeable lack of published work in that draws on both realms of research. Under strict monitoring and procedures manual control of FiO 2 can be very effective and automatic control is proven to be as effective or better than manual control is some cases. The need to take a medically justified approach to these control problems is apparent. Here we attempt to utilize the best manual control methods and draw on previous control work to synthesize a controller to replicate those results. The goal is to use the best manual control practices as a guideline for defining how the robust controller operates and therefore improve automatic control even further. The best adaptive control methods are difficult to use with non-linear systems. Using adaptive parameter estimation techniques is still a viable option but they may not be accurate enough to rely solely on their parameter estimation. Therefore a genetic algorithm method that is continuously adapting to sensor feedback is used to select between the robust controllers. 13

21 Since each robust controller is capable of closed loop stability under a range of uncertain parameters they only require a close enough estimation of the model parameters. Using this method incorporates the best of both worlds. The robust controllers will provide precise control with the safety of operating on a large range of uncertain parameters but the adaptive genetic algorithm has the flexibility to select a more appropriate controller as the patient s physiology changes. To be very clear the application of either robust control or adaptive control methods to this problem is not new or novel. This research intends to contribute the following to the literature: Adaptation of clinically tested manual control procedures into automatic control performance specifications Explore the use of a rate limiting performance specification to limit fast responses of the control effort Improve a genetic algorithm technique for real-time model parameter estimation Combine the adaptive model estimator with a robust control system in an intuitive and seamless way to provide the best possible patient care Chapter 2 discusses modeling of the premature infant s respiratory system and its response to different inputs. The genetic algorithm technique for model parameter estimation is also described. In Chapter 3 the controller synthesis is described with definitions of the performance specifications. The results of the parameter estimation tests and closed loop controller results are given in Chapter 4. The prototype that was 14

22 designed to carry out future clinical testing is illustrated in Chapter 5. Discussion of the results, limitations of the proposed design, and conclusion is given in Chapter 6. 15

23 Chapter 2: Model 2.1 Considered Models The modeling of a premature infant s SpO 2 is a difficult task. This is partly due to the wide range of birth weights and gestational ages seen in the NICU setting that represent very different levels of development. It is not easy to tailor a model for each infant and also have the ability to tune that model as the infant s health improves or degrades. The models attempt to do this by describing the relationship between the inputs and outputs as defined by the model. It has been shown that not only the FiO 2 has an effect on the patient s SpO 2 but also the HR and RR which are all easily monitored and very useful for modeling purposes. However, the problem is complicated further by changes in the infant s physiology that cannot be monitored or accounted for. An example of this is called shunting, which is when some amounts of venous blood bypasses the lungs and therefore is not transporting Oxygen to the body[26]. Sometimes this is cause by the alveoli filling with fluid which causes parts of the lung to be unventilated [27]. Circumstances such as this are difficult to monitor without invasive measures so they are difficult to account for. When such circumstances exist there is a change in the output (SpO 2 ) but no corresponding change in the measureable input (FiO 2, HR, RR). There are a few published models in this field, each with its own assumptions and limitations. Some are merely descriptive models that only attempt to map the relationship between in inputs and outputs. Others are constitutive models that use equations trying to approximate the actual physical processes of the system. Keim adopted such a model which he adapted into a non-linear relationship from a linear 16

24 respiratory system model proposed by Yu [28]. These models use an equation that describes the dissociation of oxygen into the blood, which is based on partial pressure differentials of the gases, ph, temperature, and other chemical reactions involved with respiration. This model is based on many complex physical relationships and uses safe assumptions to distill the important aspects. Since it is based on real physical interactions it is highly tunable and can be very accurate. The drawback to using such a model is that most of the tunable aspects of the model are not measureable and must be given assumed constants. It also only relates SpO 2 to FiO 2 but other measureable inputs have an effect on SpO 2. The complexity of the model, it s non-linearity and required computing power, are all drawbacks. For these reasons it wasn t selected as the main model for this research. It was, however, selected as a secondary means of verification of the models and controllers in this work as described later. Other options include the neural network model and fuzzy logic model developed in the work by Krone [23]. These two options both attempted to utilize all available feedback information to help define the model. They have the added benefit of being dynamic models. That is to say they are able to analyze sets of data and adapt the model to the data given. This is a benefit because of the desire to achieve an adaptive controller. These models were not able to achieve a level of accuracy that was sustainable over a useful length of time. The model ultimately selected was the three transfer function model that was developed by Krone. This was chosen for its use of all three measureable inputs, simplicity, accuracy in describing clinical data, and computational speed. The other 17

25 reason is for its close tie in to the genetic algorithm approach of parameter estimation for creating adaptive control. The genetic algorithm is discussed in detail in Chapter Three Transfer function model The model used to develop the controller in this work is a combination of three first order transfer functions. It uses one transfer function to describe the relationship between FiO 2 and SpO 2. The other two describe the relationship between the HR and RR to the output SpO 2 andare treated as disturbance inputs[23]. The work by Krone and Keim confirmed that the infant s response could be modeled as a first order transfer function approximation [23, 24]. These transfer functions are given in the form where G and T are the gain and time constant associated with the input x which can be HR, RR, or FiO 2. The output of each function is SpO 2. The parameters for these transfer (1) functions were defined by clinically collected data from premature infants [12]. The FiO 2 transfer function represents the relationship between the inspired oxygen and the O 2 saturation while the HR and RR functions are considered disturbance inputs which also effect the SpO 2. The model used can be seen as a block diagram in Fig

26 Fig. 2 Three transfer function model of infant response The gains and time constants that are used in this model were defined by data collected in a NICU [12, 23]. By using different sets or different combinations of these six parameters it is possible for the model to very closely match the clinical data. Estimating the best combination of parameters that most accurately describe the particular patient that it is being applied to is what is referred to as adaptive modeling. Using that information for closed loop control is considered adaptive control. Table 1 Parameter uncertainty ranges for FiO 2 transfer function Plant # Gain Time Constant Table 2 Parameter ranges for Heart Rate and Respiratory Rate HR RR Gain Time Time Constant Constant Gain (sec) (sec)

27 The entire range of uncertain parameters is broken down into six different nominal plants. This is because it is not feasible for one controller to achieve robust stability for such a large uncertainty model. The ranges of FiO 2 parameters are shown in Table 1. The ranges of HR and RR parameters are the same for each of the 6 controllers and are shown in Table 2. For both tables the nominal value used for the model is the mean of the range displayed. These ranges are used to construct uncertainty models that help define and verify the performance of the controller. As seen in Table 1 time constants of the FiO 2 ranges overlap. This is a safety net that has been designed to ease selection of which nominal plant best describes the patient. In this way a model with a time constant between two nominal plants will be accurately represented by either plant. A transfer function uncertainty model that bounds the maximum of the frequency responses of the perturbed plant describes the parametric uncertainty of the nominal plants. This model is achieved by analyzing the frequency response of the multiplicative error of each perturbed plant with respect to the nominal plant then finding one transfer function that bounds the maximum of these errors. This multiplicative uncertainty is more clearly described as the equation in frequency domain (2) where G is the nominal plant and G P is the perturbed plant. The uncertainty models shown as W HR W RR and W FIO2 in Fig. 4. will bound the magnitude of the frequency response of equation (2). An example of the possible perturbed plants and the bounding equation for nominal plant number 1 can be seen in Fig. 3. The red line is the 20

28 uncertainty model bounding all possible plants. The equation for this uncertainty model can be found along with the uncertainty for each model in Appendix A. These uncertainty models of each transfer function are used to help define the uncertainty of the entire model as in Fig. 4. where is a stable and proper transfer function with H-infinity norm less than 1. In this way all possible plants are described and can be analyzed during controller synthesis. Fig. 3. Multiplicative uncertainty for nominal plant 1 21

29 Fig. 4. Model block diagram with multiplicative uncertainty 2.3 Genetic Algorithm The genetic algorithm is a selection method used to determine which nominal plant is currently most accurate at describing the patient s response. It takes a 5 minute window of feedback information containing the three inputs and the one output and estimates which set of six parameters best describes the response. This window must be large enough to properly analyze the dynamics of the transfer function responses to inputs, but also short enough to be able to adapt to changing parameters and reduce processor time. An example of a feedback window with inputs, a simulated SpO 2, and the estimated output best guess according to the GA is shown in Fig. 5. In this example a set of model parameters and real recorded inputs were used to obtain SpO 2 data that was sent to the GA to verify its ability to estimate the correct parameters. 22

30 Fig. 5. GA window example with estimated model shown to match the feedback information The original method for using a genetic algorithm in this application was developed by Krone and is improved in this work. The estimator is sent a window of data and an array of parameters, or alleles, which are initially chosen randomly from the prescribed range. These alleles are used to create a set of models with the feedback data as input.this is done in the MATLAB environment using the lsim command and the transfer functions are represented as state space models to more easily define the initial conditions. The outputs of each model are then compared to the actual SpO 2 feedback data. Their relative fitness is described as the mean square error. The parameter set of the model with the lowest error is stored and passed on to the next generation as in evolution. The MATLAB code that is used to execute this method is shown in Appendix D. 23

31 The remaining sets of parameters are sent to a series of programs that mate certain sets of parameters and randomly mutate some parameters at a specified rate. This maintains a diversity of options, helps avoid local minimums, and also helps the system to converge on the best set of parameters. Local minimums are defined as set of parameters that are not the best descriptors of the model but changing those parameters in the right direction initially creates a larger error. The new group of alleles is then passed back to the calling function that uses them to select a controller and the alleles are saved in preparation for a new window of training data. Since new feedback data is available from the sensors every five seconds the system was designed to run on a new window of data every five seconds. Normally a genetic algorithm or any estimating algorithm would receive a stopping criteria based on error. In this case the stopping criteria is based on number of generations due to the time constraints and the need to analyze new data every five seconds. To meet this requirement the system was limited to 10 generations for each window and parallel computing techniques were used to minimize computation time. One issue to consider when estimating parameters is the occurrence of local minimums. This GA approach to parameter estimation was chosen for its ability to avoid local minimums compared to some traditional search methods. Search methods can easily find local minimums and have no means of correcting to find a best solution. An example of a local minimum for this application is the combination of 6 parameters being estimated could produce a relatively low error with a certain set of alleles but that set is not necessarily the best or correct. This is noticed more prominently when the randomly chosen initial guesses are far away from the appropriate parameters. When this occurs 24

32 the GA struggles to find the best set of parameters. Increasing the mutation rate can help combat this but that solution can also introduce unwanted oscillations in the output or cause the algorithm to continue to fluctuate dramatically when a good solution is found. To combat this phenomena a structured system of predefined guesses are systematically tested every time new data is given. The disturbance input parameters remain unchanged but each generation is tested against one other nominal plant as defined in Table 1and Table 2. This is done systematically in order to offer the genetic algorithm a better starting point if the mutation of one or multiple parameters begins to run off or if the system has stagnated at a local minimum. The decision to follow one of the predefined options is based on the same fitness calculation as the other sets of parameters. After each window of feedback data has gone through the genetic algorithm the best set of parameters is used to select a nominal plant, which will be used to select a controller, based on that nominal plant. The gain and time constant of the FiO 2 transfer function determine the nominal plant and therefore the controller that is selected. The parameter estimation process is represented infig. 6. This demonstrates the continuously updating nature of the parameter-estimating model. When the system is started an array of 6 parameter sets are randomly assigned within the appropriate range. Upon each iteration a default set of alleles is also provided for analysis along with the random. The fitness of each individual is determined in comparison to the feedback data. The sets are chosen to be mated with the roulette function then mated with the crossover function, but found in Appendix D. Randomly selected individual parameters are mutated before the process is complete and the new groups of alleles are passed back for 25

33 the next generation and the best set is saved. The window of feedback information is updated every 5 seconds so the process is started over again with the goal of finding the best set of parameters to represent the model. When the automatic control of FiO 2 is turned on the genetic algorithm estimator works the same way. The controller chooses the FiO 2 set point but the estimator receives all the feedback information and selects the controller to be used as described above. The model used to obtain the closed loop simulation results operates as seen in Fig. 7. This is also the same manner that the control system will work in a clinical setting except that the feedback information will be from an actual patient instead of being simulated with a model using known parameters. 26

34 Fig. 6. Flowchart of the parameter estimating genetic algorithm model 27

35 Model inputs & outputs are sent to the GA GA estimates the model parameters, selects the appropriate robust controller Robust controller determines FiO2 setpoint 5 seconds of life is simulated based on new FiO2 setpoint New window of feedback information is sent to GA, process is repeated Fig. 7. Parameter estimation simulations with closed loop FiO 2 control 2.4 Non-linear model Constitutive models have been developed to describe the FiO 2 SpO 2 relationship, which are based on the O 2 dissociation curve. One such model is able to more accurately model a patient s response [24]. The oxygen dissociation equation,which the non-linear model is based on,is described as (3) where P a is the partial pressure of oxygen in the artery. This equation and was adapted to model preterm infants and appropriate constants were determined in previous work and can be seen in Fig. 8[29]. The derivative is computed at regular intervals to find the gradient describing the absorption of oxygen into the bloodstream. As the hemoglobin in the blood is closer to fully saturated with O 2 molecules its affinity for Oxygen is reduces. Therefore the higher the saturation the harder it is to absorb more oxygen and the less saturated the easier it is for the blood to uptake more Oxygen. While more accurate, this model includes non-linear equations that significantly inhibit the computational speed of simulations, as well as hinders the use of established 28

36 control design techniques. It also limits the ability to validate the genetic algorithm selection process since the actual model parameters do not match the estimations process. For these reasons the non-linear model was used only as a secondary testing method for acquiring simulated closed loop results. To do this the non-linear model was simply substituted in place of the FiO 2 transfer function in the three transfer function model. In this way the disturbance inputs are still used but the FiO 2 SpO 2 relationship is more accurate. Fig. 8. Block diagram of non-linear model of FiO 2 SpO 2 relationship 29

37 Chapter 3: Control design 3.1 Performance Specifications The model was broken into six ranges of parametric uncertainty as described with the intent of designing a controller for each range. Each controller is designed to adjust the FiO 2 in a manner that emulates the best practices of an attending nurse. The protocols used in the NICU that have been clinically proven to reduce the rate of ROP in low birth weight infants are to be used a guidelines for defining the controller operation using µ- synthesis techniques to find H-infinity controllers. H-infinity control was chosen for the ease of implementation in terms the body of theoretical work that allows for clearly defined controller performance as well as robust stability and performance analysis. The synthesis methods used here and described in detail allow the strict control of key aspects of the performance which are vital to the patient s health. Other considered methods such as PID control pose unique problems such as gain tuning as patient physiology changes and possible calculation errors when sudden desaturations occur. The protocol for manual FiO 2 management is summarized as [7]: Oxygen is a drug: FiO 2 should be minimized at all times as saturation levels above 95% are potentially dangerous FiO 2 should be supplied to maintain a minimum saturation of 85% FiO 2 will not be changed frequently up and down and up again to maintain acceptable levels Weaning should be done gradually if the patient remains on the high side but can be done quickly to avoid hypoxia Do no keep an increased FiO 2 unless necessary 30

38 Small incremental adjustments of 2-5% are best, do not vary FiO 2 in large steps Alarms will not be turned off at any time. Settings must not be changed due to frequent alarms or adjustment in FiO 2. FiO 2 requirement of the infant should be documented clearly This protocol was then translated into frequency domain performance specifications that can be used to define how the controllers operate. The controller will be stabilizing the closed loop feedback control system while achieving the frequency response specifications on the control effort and error responses. As part of the µ-synthesis controller design procedure the performance weight w u (s) is defined as (4) where u max is the maximum control effort. Assuming a maximum of 100% FiO 2 and a nominal value of 21% oxygen in the air the value of u max is 79. The use of this performance weight also helps minimize the control effort, which coincides with the protocol of minimizing FiO 2 and weaning. In some cases a hospital may dictate a maximum FiO 2 that is below 100%. In those cases it would be very simple to go back and alter the value of u max to synthesize a controller that meets that requirement. According to the literature it is important to facilitate quick recovery from hypoxic events, or desaturation of the SpO 2, but also important to minimize adjustments of the FiO 2. The second performance specification helps define how the controller will respond to higher frequency events like fast desaturation, lower frequencies close to 31

39 steady state, and the crossover bandwidth which differentiates the two. This performance weight, calledw p (s), is defined as (5) where b is the bandwidth, M is the high frequency error allowable, and A is the low frequency error allowable. Using this weight defines the maximum steady state error that is acceptable, the maximum high frequency error, and the crossover bandwidth. Since the acceptable range for SpO 2 is 85%-93% if we make the controller set point 93%, the maximum steady state error is 8% SpO 2. If we want stricter control at low frequencies we can reduce the value of A to 5% or 3%. The crossover bandwidth can be set to define what frequencies will affect the control signal and what frequencies will be rejected as disturbances.this allows us to meet more of the FiO 2 management requirements by limiting the FiO 2 response to low frequency oscillations or sensor noise. This meansless changing the FiO 2 up and down and up again frequently and not keeping an increased FiO 2 setting unless necessary. The third performance weight used is the key to limiting the rate at which the FiO 2 adjustments are made. Since it is important to adjust the FiO 2 slowly and in small increments as opposed to large steps the rate limitingperformance weight is defined as (6) where is the maximum rate of change of the control effort and is the time constant of the approximate derivative and should be sufficiently small to achieve an acceptable result. The approximate derivative must be used in this case to ensure that the 32

40 performance weight is invertible when being used to synthesize the controller. For the controller synthesis, was given a value of 1.0 x This performance weight helps limit the rate at which the FiO 2 is increased or decreased. In other words it is a maximum for the first derivative of the FiO 2. This meets the protocol of varying the FiO 2 in small incremental adjustments instead of large steps. Since our hardware allows the FiO 2 to be controlled continuously, or in small time steps such as 5 seconds, and more precisely than a nurse, the FiO 2 can be controlled in a more optimal way in terms of the infant s health. A nurse may only be able to determine a resolution of 1-2% FiO 2 and might only be able to adjust the FiO 2 every few minutes. This performance specification helps the controller match the same result but with slow continuous adjustments. For controller synthesis the performance weights and uncertainty models are applied to the three transfer function model with the RR and HR aspects re-arranged and labeled as disturbances. This can be seen in Fig. 9 along with the controller K which is to be determined. In this figure the W blocks for HR, RR and FIO 2 represent the uncertainty model that is multiplied by the block which is some transfer function with H <1 and is stable and proper. 33

41 Fig. 9 Block Diagram of model, performance weights, and uncertainty model 3.2 H-infinity synthesis The controller K is to be synthesized such that the closed loop system is stable and the H-infinity norm between the inputs and outputs of the matrix in Fig. 9are less than one. If this condition is met then the resulting controller is said to have robust performance, which means that for the defined range of uncertain model parameters the close loop control meets the desired performance specification. The µ-synthesis process to find the actual controllers was done using the hinfsyn.m command in MATLAB.This command uses a mixed sensitivity approach, which minimizes the cost function 34

42 containing the three performance weights from above and this cost function must be less than one to meet performance specifications as seen in equation (7)[30]. (7) where S is the sensitivity transfer function of the closed loop system, and T is the complementary sensitivity function. The generalized plant transfer function matrix, P, that represents the system in Fig. 9and is used for the controller synthesis process is shown in Equation (8). Once this synthesis process was completed for each of the 6 controllers verification of their performance is completed to ensure the performance specifications were met. 35

43 (8) 3.3 Performance Analysis Verification was done to ensure that each controller, K, found by the µ-synthesis technique meets robust stability requirements and the performance specifications. To ensure the controllers are nominally stable with no plant uncertainty and there must be no right-half plane poles for the nominal closed loop plant. One performance check was to ensure the value of Equation (7) was less than one.another method includes analyzing the N matrix. This is found through the lower fractional transformation of the P matrix described in Equation (8) and the controller K. The inputs and outputs of the N matrix follow the labeling scheme of Fig. 9. and are shown as 36

44 (9) Nominal performance checks that the nominal plant is controlled according to the performance specifications. This is done by ensuring that the H-infinity norm of the lower partition of the N matrix is less than one, Nominal performance: (10) Robust stability and robust performance ensures that the closed loop system is stable and meets the given performance requirements over the given range of uncertain parameters.the necessary condition for robust stability is given as, Robust Stability: (11) The final test is robust performance, which ensures performance specifications are met for all uncertain plants. This is confirmed if the N matrix has a maximum singular value less than one. Robust performance: (12) The performance criteria used to define the performance weights in some cases was different than the performance weights used to calculate the performance checks. 37

45 This was done in instances when the optimization process for the control synthesis did not find a controller that met the original performance specifications. Modified performance weights were applied in order to ultimately find the desired performance. In all cases the controllers were able to meet the original criteria required but fine-tuning the synthesis process was required. 38

46 Chapter 4: Results 4.1 Controller Synthesis results The synthesis of the six controllers was completed as described in the previous chapter. To verify that each controller, K, met the performance criteria that were initially specified, Equation (7) was checked and confirmed. The criteria outlined in Equations (10) - (12) were checked for all six instances but did not prove robust stability and robust performance for all disturbance inputs. When the disturbance input uncertainty models W HR and W RR are reduced to zero the system does meet the robust performance criteria. The performance criteria results for the controllers without considering the disturbance uncertainty are show in Table 3. A full description of the controller transfer functions and an example bode magnitude plot can be found in Appendix B. Plant # Table 3 Performance and Stability verification values Nominal Stability (# RHP) Nominal Performance Robust Stability Robust Performance It is also useful to see the actual controller response to understand the benefit of each performance specification. To ensure that the controllers were operating in a desired manner and that the rate limiting performance specification was working as desired a simple step command was given to the system. In Fig. 10 the control signal 39

47 response to a step input is shown when a controller is synthesized without using the rate limiting performance specification. It should be noted that the control effort instantly jumps up to 12% FiO 2 then quickly drops to a small steady state value. This means a large fluctuation in FiO 2 just to get a quick recovery that is most likely unnecessary, and possibly harmful to the baby s health. Fig. 10. Closed loop step response of controller without rate limiting performance specification Compare the results in Fig. 10. to the response of a controller that was synthesized with the addition of the rate limiting performance specification, which can be seen in Fig. 40

48 11. This response varies greatly in terms of the maximum control effort achieved, as well as the rate in which the FiO 2 was varied. In both cases the same steady state error and same steady state control effort was achieved. In the second case the system took slightly longer to achieve steady state. That is acceptable according to our goals and from the perspective of the patient s health, potentially very beneficial. Fig. 11. Response of a controller synthesized using the rate limiting specification 41

49 4.2Open loop parameter estimation This approach for controlling the premature infant s SpO 2 is based on the ability to accurately estimate the model parameters and select the appropriate robust controller.it does this by matching the best set of parameters to one of the six nominal plants as described previously. For this approach to work the algorithm must be able to pick the best controller reliably and the controller must function properly. The effectiveness of the genetic algorithm in selecting the appropriate controller was determined using a series of tests. Since it is impossible to verify the actual model parameters of clinically collected data, a way of creating simulated SpO 2 to validate the GA was developed. Hour long strings of clinically collected FiO 2, HR, and RR data were selected as the inputs to the model. Model parameters were then chosen for the three transfer functions and the input data was used to obtain a simulated SpO 2 response. This data can now be used to verify the efficacy of the genetic algorithm because the inputs are realistic and the model parameters are known. The input data and SpO 2 are given to the genetic algorithm using its updating window approach to parameter estimation. As the window moves up every 5 seconds a new set of information is provided to the genetic algorithm. It runs 10 generations and provides a set of 6 model parameters, gain and time constant of the transfer functions that describe each input,that correspond to the lowest error. Since the FiO 2 gain and time constant are used to select which controller is appropriate, it is important that their parameter estimation is accurate. This selection is given in the form of the nominal plant number, 1-6, where each plant is described previously in Table 1 and Table 2. 42

50 4.2.1 Constant parameter estimation Simulations were arranged with parameters falling in the range of each of the 6 controllers to test the ability to select each one accurately. The three transfer function model was used for this simulation so that the known model parameters could be compared to the estimated ones. The data was prepared with a time step of 0.1 seconds to match the time step of the GA and ensure stability. The parameter estimation for a case where controller number one is the correct solution is shown in Fig. 12. The red horizontal lines indicate the correct parameter that is to be estimated and the blue represents the best guess at each iteration. This simulation was done for 200 iterations of the genetic algorithm, which represents a little over 16 minutes of life for the baby. 43

51 Fig. 12. Controller 1 parameter estimation simulation The result in Fig. 12 shows that the FiO 2 transfer function gain is estimated to an acceptable level of accuracy by the first iterations of the algorithm and maintains a close approximation the entirety of the simulation. In terms of selecting the best controller the gain portion of the algorithm is only responsible for deciding if the gain is above or below a value of 3. In most cases the gain estimation is accurately estimated within 20 iterations. The time constant was randomly assigned an initial guess that was substantially further away from the correct parameter but the result shows how it continues to improve over the next 10 minutes of simulated life and eventually finds the 44

52 correct solution by the 70 th iteration. The error calculated for each moving window of data is shown in Fig. 13. which illustrates the larger error associated with the poor parameter estimation near the end of the simulation in Fig. 12. A representation of which controller the algorithm selected during this simulation is show in Fig. 14. As can be seen, it ended up on the correct controller except for one time step when it estimated a different plant but then returned to the right choice on the next iteration. This happens occasionally and is a built in aspect of the genetic algorithm. Mutations such as this allow exploration of other options and provide opportunities to jump out of local minimums. The random selection of a different controller that could alter the optimal control of FiO 2 is easily remedied by a requirement for a sustained choice for a defined amount of time or averaging which would eliminate outliers while still allowing the GA the necessary freedom to make those mutations. Fig. 13. Error used to assign fitness of parameter sets during simulations 45

53 Fig. 14. Yellow line represents the correct choice, red is simulation result The other 6 controllers exhibit very similar results when the correct model parameters are close to the nominal plant parameters. Occasionally when the correct parameters are between two plants the best guess will fluctuate between those two controllers. Since the ranges that the error models and controllers were designed with overlap, this oscillation or indecision is not a problem in terms of robust stability. If the correct model parameters lie in the overlap range of the robust controllers then both choices will operate with robust performance. The result when the correct time constant is between controllers 2 and 3 is shown in Fig

54 Fig. 15. Result when the correct parameters are between controllers Changing parameter estimation The genetic algorithm must also be able to adapt to changing parameters, not just select one set reliably. To test its ability to adapt to changing parameters and simulate its tracking of physiological changes in the patient another test was devised. An hour-long set of inputs was fed into the three transfer function model as before but at specified intervals the time constant of the FiO 2 transfer function was changed. The window of inputs and response then represented changing parameters that the GA must estimate. The first test was to ensure it could estimate increasing time constants. The FiO 2 gain remained constant while the time constant increased from 80 to 145 and then 185. This was done over 50 minutes and the time constant increased every 1000 seconds. The results of this test can be seen in Fig. 16 and Fig

55 Fig. 16. Estimating varying time constants test Fig. 17. Controller selection for varying time constant test 48

56 The gain selection in this result is very accurate as can be observed in the top left plot of Fig. 16. The time constant estimation in the top right doesn t have perfect accuracy but it does track the changing parameters. After it drifts away from the best solution and occasionally selects the wrong nominal plant, it is able to correct itself and return to a better solution. The time constant selection appears to anticipate the change in parameters but this is a result of the window of past data being used to estimate. This simulation, Fig. 16 -Fig. 17, for showing the ability to adapt to changing model parameters is a little unrealistic compared to what you might see in the physiology of a real baby. Such drastic increases in the time constant of FiO 2 response would not occur two times in one hour. This was done merely to illustrate the functionality of the system in as simple a way as possible. If it can estimate large parameter changes quickly it will be able to detect small gradual changes that much easier. Unpredictable changes in patient physiology that cause the time constant of the response to shorten, although rare, can happen. To show the ability to switch between the remaining three controllers and also the ability to estimate time constants that are decreasing another simulation was developed. The parameter estimations for the test that changes from plant 6 to 5 to 4 can be seen in Fig. 18 and the controller selection can be seen in Fig

57 Fig. 18. Estimating varying time constants test 2 Fig. 19. Controller selection, varying parameters test 2 50

58 As the patient grows and gains weight the volume of red blood cells increases, muscle mass increases, and other physical characteristics change. This could lead to changes in the system gain. A test to illustrate the model s ability to adapt to changing system gains is seen in Fig. 20 and Fig. 21. The estimator is able to track the changing model parameters and select the appropriate controllers. Fig. 20. Changing gain parameter estimation 51

59 Fig. 21 Changing gain controller selection Parameter estimation with real data Other means of verifying the viability of the genetic algorithm were employed to ensure that it has the ability to operate in a clinical environment with real patients. To do this the same algorithms were applied to sets of real inputs as before. However, the SpO 2 data was not simulated this time. Since the genetic algorithm is running on real data these simulations will give insight to real time operation but verification of parameter selection is impossible. Hour long sections of data with a high number of FiO 2 changes were selected to ensure adequate excitation of the transfer functions. The raw data was sent to the genetic algorithm with a window of 5 minutes, which proved to give the most consistent results. An abbreviated example of these resultsis shown in Fig. 22 so that more detail can be 52

60 seen. There is no red line to indicate the correct parameters since they are unknown. It was noted that the magnitude of the mean square error in this case was on the order of 10 to 50 times higher than the results above where the SpO 2 was simulated. This is to be expected since the simulated SpO 2 was created from the model that was used to do the estimating. The model will never match the actual patient response perfectly since it is only simulating a very simplified representation of the patient. The controllers selected during this test are shown infig. 23. and the input data is shown in Fig. 24. Fig. 22. Raw data simulation with unknown model parameters 53

61 Fig. 23. Raw data simulation controller selection Fig. 24. Input data from clinical setting used in GA testing 54

62 4.3Closed loop simulations The goal of this adaptive parameter estimation technique is to ultimately select a controller that will effectively control the SpO 2 of the patient. It has already been demonstrated that the algorithm can select the correct controller and that the controllers meet the performance criteria. Here it will all be put together to demonstrate the selection and closed loop control at the same time. Disturbance input data was derived from actual patient recordings as before. The FiO 2 was initially taken from patient readings as well but after the first iteration of the genetic algorithm a controller is selected which dictates the FiO 2 setpoint on the fly. Model parameters were given which fall under controller number 4 and for the entirety of the test the GA did select number 4. The parameter estimation can be seen in Fig. 25. More importantly the closed loop control history is shown infig. 26. where the values shown are normalized so that the desired control set point and the nominal FiO 2 are zero. The controller sets the FiO 2 and successfully maintains the SpO 2 within 1% despite the disturbance inputs. Notice the smooth adjustments in FiO 2 that limit overshoot. This control simulation confirms the operation of the controller according to the specifications and its ability totrack a small steady state error. The same test was done with parameters that fall within each nominal plant. In all cases a similar result was seen. When automatic control takes over in these simulations the SpO 2 is regulated as desired. 55

63 Fig. 25. Closed loop parameter estimation Fig. 26. Closed loop control test SpO 2 and control effort 56

64 A second closed loop simulation was done for a longer time period to illustrate the controller s performance and ability to maintain a healthy SpO2with adverse disturbance inputs. The control effort, or FiO2, and the SpO2 response can be seen in Fig. 27. The corresponding disturbances which are creating the small changes in SpO2 can be seen in Fig. 28. It should be noted that the resolution of the feedback SpO2 in these simulations is not commensurate with the existing sensor technology implemented in the hospital. These simulations provide high resolution data because they are based on calculations. The minute changes in SpO2 seen in Fig. 27. would not be seen in a clinical setting and the control effort would not respond to those small changes as minutely. Fig. 27. Modeled closed loop results with real disturbance inputs from data 57

65 Fig. 28. Real disturbance input data used 58

66 Chapter 5: Prototype design and testing A prototype device for carrying out automatic control of the FiO 2 in a clinical setting was needed. Previous clinical trials have not published designs for their means of controlling FiO 2, only describing their necessary attributes. Presented here is an upgraded design for retrofitting a standard manual FiO 2 control valve into one that safely accepts automatic control. The design consists of an IV stand for mounting, a blend valve which is to be manipulated, a motor for control, a mounting system, and a microcontroller for accepting sensor feedback and making calculations. An assembly of these various parts is shown in Fig. 29. to help illustrate the components described below. Fig. 29. NEO2 Blend valve retrofit design assembly drawing. 59

67 5.1 Device Performance Specifications The performance specifications of this device are intended to match the needs of easy operation in the hospital, safety for the patient, and simplicity of design for reliability. Specifications for the overall size of the device were outlined by University Women s and Children s Hospital and easily met. As a whole it could not be bulky, protrude far beyond the face of the blender, and it must function similarly to the unmodified product. It was also mandated that the functionality of the blender or other devices not be obstructed in any way, meaning manual control must be possible at all times. The overall concept of retro-fitting an existing O 2 mixing device eliminates unnecessary complexity and provides the same level of accuracy, reliability, and safety of the current industry standard in terms of the gas mixing mechanism. Since these devices are already in place in every NICU, there is no need to re-invent the wheel by designing a new O 2 mixer. The modification to the blend valve was influenced by the need to safely attach a motor for automatic control without hindering the normal operation by the user. Operation of the device mimics the original blender in almost every way. There is a toggle switch for when manual control is desired, and a sensing element is in place so that the automatic control shuts off when a manual input is detected. The custom knob and indicator dial were designed to mimic the original appearance and operation of the blend valve prior to modifications. With the motor tucked safely away from the operator, the manual control knob and indicator dial allow normal operation of the blend valve when automatic control is turned off. Even when the device is shutdown and unpowered, 60

68 the manual control aspect of the blend valve exists exactly as before the modifications were made. This is important for the safety of the patient as well as for ease of use by the attending nurse or doctor. When automatic control is turned on, an indicator light is turned on to assure that the electronics are functioning properly. Whenever the microcontroller box is powered on but automatic control is not enabled due to operator choice or malfunction, this indicator light flashes to show an error. These lights are located on the front panel of the microcontroller housing in Fig. 30. Other performance specifications include the ability for the electronics to effectively communicate with the pulse oximeter, oxygen analyzer, and Spacelabs system to get proper feedback. This is done with serial cables, and communication with the attached laptop is possible with an Ethernet connection. It was also specified that a calibration procedure be in place to ensure accurate feedback and O 2 control by the motor. In order to achieve this, a dialogue box provides instructions and forces the user to calibrate the device upon startup. There is also the option to re-calibrate during use if it is observed that the FiO 2 read by an oxygen analyzer does not match the input command. The computational capabilities and computer hardware of the device have already been designed Krone and Keim and were not altered. The hardware and software allows feedback from FiO 2, SpO 2, HR, and RR sensors, which are connected via serial cable. It has the ability to control two motors so that a flow control algorithm can be added. It also has a vibration motor to physically stimulate the patient if HR or RR becomes irregular [23]. 61

69 Some novel programming improvements have been made to account for the new physical aspects of the device. New wiring was added to the FPGA module to support the automatic control switch and lighting. The new DC servo motor hardware also required changes in the software. Control logic that will power off the motor when a manual intervention is recognized was added. Upgrades to the user interface for easier operation and recording of data were completed. Fig. 30. Control panel and hardware box 5.2 Drawings and Final Design Krone and Keim completed the first iteration of the prototype. This design achieved the desired ability to automatically control the FiO 2 by retrofitting the NEO2 blend value but did not meet size or safety and operation requirements. This original design can be seen in Fig. 29. The drawbacks included a lack of easy manual control, bulky overall size, a large exposed belt system, and a motor gearbox that did not turn freely when unpowered. The motor chosen, the belt system, and lack of knob meant there was no manual control override system that is crucial to limiting risk to the patient. 62

70 These limitations were revised in the second iteration of the prototype. The distance that the device protrudes from the face of the blend valve was reduced by ½. The large motor and high friction gearbox were replaced with a smaller one with no gearbox that allows it to rotate freely when unpowered. A miter gear system replaced the belt drive to accommodate a manual control knob. A replacement indicator for the dial face of the blend valve was 3D printed along with the new manual control knob that mimics the original hardware. The improved design can be seen in Fig. 32. The design was fabricated with aluminum panels, cut and drilled to mount the motor shaft. Pillow blocks were used to support the shaft. A nylon miter gear system transmits the motor s power. A helical beam shaft coupler attached the shaft to the blend valve and allows for small misalignments during operation. Detailed engineering drawings for these parts are illustrated in Appendix C. 63

71 Fig. 31. Original prototype design, circa 2011 Fig. 32. New prototype design completed 64